Insider Brief:
- A study from the Beijing University of Technology, Beijing Science and Technology Project Manager Management Corporation Ltd, and the University of Nottingham introduces a hybrid quantum–classical model that integrates variational quantum algorithms with 3D CNNs to improve the accuracy and stability of Laser Speckle Contrast Imaging (LSCI) for blood flow imaging.
- By replacing traditional pooling layers with variational quantum circuits, the model retains spatial and temporal data, which reduces information loss and improves predictive performance.
- Initial tests using synthetic tissue models showed that the quantum–classical model outperformed classical methods by up to 26.1% in prediction accuracy.
- Although further in vivo testing is needed, this model could extend to other imaging modalities like MRI and CT, potentially setting a new standard for precision diagnostics.
Accurate blood flow imaging is essential for diagnosing and managing a wide range of health conditions, from cerebral circulation issues to vascular complications in diabetes. Yet, achieving high precision in these measurements is no trivial feat. A recent study published in Nature’s Scientific Reports proposes a hybrid quantum–classical framework for Laser Speckle Contrast Imaging (LSCI), an optical method widely used to assess blood flow dynamics. According to the study, a collaboration between the Beijing University of Technology, Beijing Science and Technology Project Manager Management Corporation Ltd, and the University of Nottingham, this framework integrates variational quantum algorithms with 3D convolutional neural networks to provide improvements in prediction accuracy and model stability for blood flow imaging tasks.
The Limitations of Traditional Laser Speckle Imaging
LSCI technology, known for its ability to visualize blood flow without requiring contrast agents, has long been used in medical fields ranging from cerebral and retinal assessments to trauma and burn evaluations. However, while traditional LSCI provides valuable insights, it remains largely qualitative, as it struggles with precise blood flow measurements due to inherent limitations. As the study points out, LSCI often relies on approximate models that fall short in accurately capturing quantitative data, especially when faced with complexities such as static scatterers—non-moving particles that can interfere with imaging clarity by scattering light in unpredictable ways—and variable speckle sizes.
To address these challenges, machine learning models, especially classical 3D CNNs, have been integrated into LSCI to take on the spatiotemporal data. While effective at improving accuracy, these models often use downsampling techniques, which, according to the study, can result in substantial information loss. Downsampling methods are used to reduce data resolution or size for convenience, but they often lead to a loss of detail in the process. This limitation reduces the model’s ability to fully incorporate the intricate spatial and temporal patterns in LSCI data, and ultimately compromise any predictive performance.
Quantum Algorithms as a Solution to Information Loss
In this study, the researchers introduce a quantum–classical hybrid model that addresses the information loss seen in conventional 3D CNNs. Instead of using the standard 3D global pooling layer, the layer which compresses feature maps into singular values per channel, the hybrid model replaces it with a variational quantum circuit. This VQC preserves the spatial and temporal relationships within the data to preserve the model’s ability to make accurate predictions.
As noted in the study, VQAs allow the model to optimize a parameterized quantum circuit by using classical computation, making them especially suitable for NISQ environments. This framework avoids the pitfalls of overfitting often seen in classical models, thanks to the efficient data encoding and expressivity of VQCs. Unlike traditional pooling, VQCs make it so the model can use the entire feature map, retaining the spatiotemporal information that would otherwise be lost.
To test their hybrid model, the researchers conducted experiments on a dataset of speckle data from a specially engineered tissue phantom—synthetic model designed to mimic the optical properties of human tissue—that simulates blood flow under various controlled speeds. Through cross-validation, the hybrid model demonstrated up to a 14.8% improvement in mean squared error and a 26.1% improvement in mean absolute percentage error as compared to classical 3D CNNs.
According to the study, this improved performance is attributed to the quantum model’s ability to capture complex patterns within LSCI data, providing more stable learning curves and higher prediction accuracy. Interestingly, the quantum models also excelled in generalizing to new, unseen data—a notable factor in medical applications where model reliability on diverse patient datasets is essential.
Remaining Challenges and Future Directions
While the study demonstrates improvements in prediction accuracy for blood flow imaging, certain limitations remain. As noted by the researchers, the model’s current validation is based solely on experimental setups using tissue phantoms, which simulate human tissue but do not capture the full complexity of live biological systems. Future research will need to expand these validations through in vivo testing to confirm the model’s clinical applicability.
Additionally, due to computational constraints, the researchers could only use a limited number of image frames for training, which may impact the model’s ability to capture the full scope of blood flow dynamics. Scaling up frame counts and exploring more resilient quantum hardware are other variables that may positively impact the model’s performance as quantum processing capabilities mature.
However, the results of this study are an important contribution in the larger scheme of adapting quantum machine learning to medical imaging. Through more accurate blood flow assessments, this hybrid quantum–classical framework has the potential to advance different diagnostic areas, from monitoring diabetic foot ulcers to evaluating cerebral blood flow. As the researchers note, the model’s ability to retain full feature maps from LSCI data means it could be adapted for other medical imaging modalities that rely on volumetric data, such as MRI and CT scans.
Toward Clinical Precision: Quantum’s Role in Medical Diagnostics
Future research will focus on validating this framework in vivo, expanding beyond experimental setups. While current quantum computing hardware imposes some constraints, ongoing developments in quantum processing could make these models even more accurate and accessible for clinical use.
The quantum–classical hybrid model’s ability to retain essential spatiotemporal information makes it a valuable tool for not only for LSCI, but potentially for other applications that rely on both predictive accuracy and generalization across diverse datasets. As quantum technology progresses, models like these could become foundational for precise, non-invasive diagnostics.
Contributing authors on the study include YiXiong Chen, WeiLu Han, GuangYu Bin, ShuiCai Wu, Stephen Peter Morgan, and Shen Sun.
